Comprehensibility

As with the Nintendo data, binary classification tasks were given to investigate readability. The tree produced for the sign thank is shown in Figure 6.30. The sign thank was chosen specifically to make comparisons with the results for the Nintendo sign data shown in Figure 6.27 possible.

The rules are much simpler. It makes use of the additional channels. First, it uses the roll of the hand. The first rule says: if the right hand does not roll to a palm up position early in the sign, then it can not be the sign thank. This is a logical first-cut: there are very few signs that begin in a palm-up position and it eliminates about 92 per cent of the training instances.

The second rule checks to see if the hand is moved horizontally and laterally towards the body. If it has either of these characteristics, then it can not be the sign thank.

The third rule checks the ring finger (a good indication of the state of the palm). If the ring finger is, on average, very close to unbent (as would be the case for an open-palm sign like thank) and the left hand doesn't move horizontally at all, then it is the sign for thank. The ``at all'' comes from looking at the event index. The bounds on event 6 are quite large, for example, the midtime is anywhere from 0.0 to 0.82, and the duration is anything from 0.04 to 1.0. These values are expressed in relative terms: in other words a duration of 1.0 would mean the length of the whole sign.

Figure 6.30: A decision list produced by TClass on the Flock sign data for the sign thank.

It's obvious that this rule is clearer than the rule shown in Figure 6.27. It uses 20 per cent fewer events and rules and uses the additional channels in a way that produces a much more intuitive description.

Figure 6.31: Events referred to in Figure 6.27 with bounds.